Dmytro Hrybov

Hi, you can also call me Dima. I love machine learning and electronic music. I've spent eight years building ML models for all sorts of fun applications, mostly computer vision. Currently in San Francisco.

Now

  • Since Aug 2026

    Founder resident at Entrepreneur First (The Bridge, S26). I'm benchmarking frontier VLMs on physical-world tasks, using real robot episodes rebuilt as replay-verified sim environments. Earlier on, I fine-tuned MiniMax H3, a 33B video diffusion model, into a robot policy that predicts both video and actions, trained on DROID.

Before

  1. Apr–Aug 2026

    ModelRoomCo-founder & CTO

    We built agents that plan and produce multi-shot videos, plus a VLM harness that checks every shot against the prompt and against the rest of the video. I wrote the core and led four engineers.

  2. 2025–26

    GoogleSoftware engineer, Kaggle Research

    I reproduced MMLU, FACTS Grounding and other major LLM benchmarks and hosted them in the open on Kaggle Benchmarks. I also built the agentic pipelines behind the Data Science Agent in Colab Enterprise, and led the GPU region expansion that brought B200, H200 and H100 capacity to more GCP regions.

  3. 2023–25

    PromatonSenior machine learning researcher

    3D machine learning on dental scans: reconstructing and deforming meshes, segmenting and generating point clouds, estimating pose. I ran several new projects end to end, from system design and data collection through research to deployment.

  4. 2020–23

    SurgalignMachine learning engineer

    I trained neural networks on 2D and 3D scans for spine surgery: spinal cord instance segmentation, pedicle screw placement, 3D patient orientation. I led a five-person team annotating CT and O-arm volumes, and rebuilt the training pipeline to get 240% more throughput out of it.

  5. 2019–20

    TCL Research EuropeAI engineer

    Image enhancement research for TVs and phones, with HDRNet, GANs and transformers alongside detection and segmentation. I improved a human-body segmentation network by 4.09% mean IoU and co-authored the NTIRE 2020 challenge on video quality mapping (CVPR Workshops).

Also

  • 2022–23

    Real-time generative installations for science and art events: a StyleGAN3 that turned live audio into video, and an image-to-image SDXL Turbo I optimized for speed. Both ran at over 30 fps.

  • 2016–20

    BSc in computer science (machine learning) at the Polish-Japanese Academy of Information Technology in Warsaw.

  • I speak English, Ukrainian, Polish and Russian fluently.